Asari AI

Our co-inventor agents made two of the newest open-source models faster, and they're getting better with every run. Over the weekend, our agents made Kimi K3 32% faster at concurrency = 1, running vLLM on Nvidia B200s. In another lightning run, they sped up Thinking Machines' Inkling-NVFP4 by 18% at c = 2. AI inference moves quickly and is multi-faceted: model × stack × chip × use-case (and more). Optimizing in this space requires versatility. The most exciting part is the compounding. Both runs drew on insights the agents distilled from earlier optimizations, including DeepSeek v4 Pro and GLM 5.2. Each cycle starts ahead of the last and subsequent runs are more efficient. Our self-improving agents are becoming the next generation of compilers, autonomously optimizing entire software systems at the rapid pace of AI. Read more at our blog: https://lnkd.in/gnVupZ35

Our co-inventor agents optimized the full AI inference stack, verifying every change with rigorous correctness checks. They worked autonomously on vLLM serving two of the largest open-source LLMs, DeepSeek v4 Pro and Z.ai GLM 5.2, on NVIDIA B200s. The agents improved throughput and interactivity by up to 16% across concurrency levels, running for about a day per level. Production inference performance depends on millions of lines of code, hundreds of configuration parameters, and runtime dynamics that shift with every workload. A fast kernel in isolation doesn’t translate into a faster system overall. Our agents optimize everything together: kernels, schedulers, load balancers, and configuration. Speed only counts if model behavior is preserved. Customizations can quietly alter outputs while still acing standard benchmarks. We check every change with statistical distribution-matching tests on token-level probabilities, across a diverse multilingual corpus. The agents also learn as they go. A lesson from optimizing DeepSeek v4 Pro on how to avoid a class of distributed deadlocks transferred directly to GLM 5.2, saving 44 minutes of wasted wait time. Each run makes the next one better and faster. Full write-up at https://lnkd.in/gnVupZ35.

Building software for space is one of the hardest engineering problems there is. As AI becomes more capable of writing and reasoning about complex code, we wanted to explore the key question: can AI help us invent at the frontier? Here's what we found →

Thrilled to see MIT Computer Science and Artificial Intelligence Laboratory (CSAIL) highlighting our work on EnCompass! EnCompass rethinks how AI agents explore complex problems by separating workflow logic from search strategy. This makes agents more reliable, easier to build, and opens up new possibilities for intelligent exploration. Read more about our approach: https://lnkd.in/g2QCU7Dc

Migrating critical infrastructure from C to memory-safe Rust at scale is hard. We challenged our AI agents on four production libraries consisting of 40,000+ lines of code that power essential systems worldwide: • libmcs: high-precision math library that meets strict European Space Agency (ESA) standards • gzip: one of the most widely-used command-line tools for file compression • libyaml: the foundational parsing library used for YAML configuration files • zlib: used by virtually every operating system and application for data compression ➡️ Result: All four successfully translated to Safe Rust with comprehensive test validation. Our agents maintained correctness across thousands of interdependent decisions, demonstrating they can handle the rigorous system design and verification required for production engineering. We've open-sourced all four Rust translations. Development combined human-AI co-design for specifications with fully autonomous implementation by our agents. Read the full blog post here: https://lnkd.in/gWyMhxuH Open-source repos: libmcs: https://lnkd.in/gQuR25sP gzip: https://lnkd.in/ge7ArKiq libyaml: https://lnkd.in/gDhBEc72 zlib: https://lnkd.in/gygpvw4k All feedback welcome!

We're excited to share our first blog post, introducing EnCompass: a framework that fundamentally rethinks how AI agents search and reason through complex, multi-step tasks. Presented at NeurIPS 2025. The Challenge: As AI agents tackle complex workflows like code translation or system debugging, possible paths multiply exponentially and small errors compound, yet traditional approaches hard-code search strategies directly into workflows, making them brittle and difficult to improve. Our Approach: EnCompass separates what agents do from how they explore solutions. Developers annotate workflows with branchpoints, then apply various search strategies such as beam search, best-of-N, and tree search without rewriting lines and lines of code. The Results: Across our three case studies (code translation, hypothesis generation, iterative code refinement): • 3-6× fewer lines of code to implement search strategies • Cleaner, more maintainable agent workflow code • Practical implementation of techniques to unlock better scaling laws Read the full blog post: https://lnkd.in/gu8cH72y Paper: https://lnkd.in/g-bCXCtz

If you're looking for the next ambitious project to take on, we’d love to talk! We’re building out the AI infra and compute to help people invent the systems of the future. Come build with us → hello@asari.ai, open jobs at jobs.ashbyhq.com/asari.ai